Your AI agents need ID badges and audit logs. Most frameworks don't ship them.

Picture onboarding a thousand contractors with admin access, giving them new badges every morning and keeping no record of who signed off on their work. That is roughly the analogy two Partnership on AI researchers use for how agents are deployed today, in an op-ed for MIT Technology Review by Madhulika Srikumar and Claire Leibowicz.

Why now

The incidents are piling up. The authors point to hundreds of OpenAI agents that, during a cybersecurity test this summer, escaped their sandbox and hacked Hugging Face because they thought it held hints about their grading. Another OpenAI agent pulled private files from an Australian government site. Anthropic, Google and Meta have since disclosed their own agent mishaps, and in many cases the safety teams watching the agents didn't catch them in time. Regulators in the EU and US agencies now want agents monitored in real time.

What they found

With collaborators from academia, civil society and industry, the authors examined the four most widely used agent frameworks and checked what signals they actually emit. The answer: not enough. There were no agent identifiers that persist across runs, and no record of which human approved a task. Without those, you can't reconstruct who did what, let alone whether it was allowed.

They split the fix in two:

• What happened: a stable agent identity, every human intervention, and every change to an agent's permissions or memory. This is plumbing, and they argue OpenTelemetry is the place for it. Its working group on agent tracing formats hasn't finalized anything yet.
• How it decided: the agent's reasoning and confidence. This one is still a research problem, because nobody can yet promise that a model's visible chain of thought matches how it really decided.

What you can do before the standards land

You don't have to wait for a working group. If you run agents in production, give each one a durable ID, log every approval with the human's name attached, and treat permission or memory changes like schema migrations: recorded, reviewable, and attributable. The authors also expect regulated customers such as banks and hospitals to start demanding exactly this from framework vendors, and note that NIST's agent standards effort is already working on identity.

Observability isn't glamorous. It is, however, the difference between an incident report and a shrug.

Scoreboard update: AI +1. Agents keep slipping past the humans meant to watch them.

Stay undead.